AMD: Volcanic Islands R1100/1200 (8***/9*** series) Speculation/ Rumour Thread

Those specs from 3DCenter.org are not final or 100% correct.
They could be correct, but they have not to.

What does that even mean?

Quad-Frontend. As far as I know, GCNs-Frontend is already scalable. Instead of two rasterizers and tesselators, we might get 4 rasterizers and 4 tesselators.
 
Tahiti is "quad engine".
Was 2 at launch:
architecture.jpg
 
Depends on what you term as engines - that's talking about geometry engines. In terms or shader architecture, why do you think Tahiti scales SKU's in 4 CU increments vs 2 in other products? The architecture is scalable in multiple vectors.
 
Tahiti is "quad engine".
Probably depends on what engine you are talking about. :LOL:

Edit: Too slow, was looking up the GCN presentation what was called an engine there.
Depends on what you term as engines - that's talking about geometry engines. In terms or shader architecture, why do you think Tahiti scales SKU's in 4 CU increments vs 2 in other products? The architecture is scalable in multiple vectors.
So you meant shader engines. And isn't Pitcairn Pro not also using just 16 CUs?
 
So you meant 4 rasterizers (which also seems to be what 3dcenter is speculating about) in tahiti, though the drawing only got 2 for simplicity?

(for Hawaii I would assume it is more important to speed up the geometry setup than the rasterizers)
 
Yes, forgot, Pitcairn is the same org as Tahiti. Verde and Bonaire are 2.
Which is kind of strange as we were told, that Pitcairn uses 6 CU groups (2 x 4 CUs + 4 x 3 CUs). How is that split? Or can CUs from the same group belong to different shader engines?
 
Pitcairn has 4 CU groups as has Tahiti.
You mean shader engines as Dave and me just agreed :LOL:.
With the CU groups I meant the groups of CUs sharing the scalar L1 (kcache for some of the graphics guys) and instruction cache. And CapeVerde has reportedly a setup of 4+3+3 CUs (that means it has three L1-I and three sL1-D caches), while Pitcairn just doubles on this to arrive at 20 CUs with 6 groups (2 groups of 4 CUs and 4 groups of 3 CUs). The die shot also implies that one has 10 CUs per side opposed to each other. I guess it would be weird to distribute 5 groups of 4 CUs in such a way. Alternative would be AMD screwed us all over with their presentation of the GCN architecture and the shared instruction and scalar data cache. You have the choice. ;)

The question is if CUs from a single physical group can be distributed over different shader engines. Guess there are corner cases where this isn't exactly optimal, but who knows? Maybe someone in your company?
Or CapeVerde and Pitcairn use significantly smaller CU groups. CapeVerde would have 4 in total (distributed to 2 shader engines), so likely one group with 3 CUs plus one group with 2 CUs forms a shader engine, same as with Pitcairn. That would resolve this, too.
 
Last edited by a moderator:
You mean shader engines :LOL:.
With the CU groups I meant the groups of CUs sharing the scalar L1 (kcache for some of the graphics guys) and instruction cache.
I only used the terminology you used above. Pitcairn and Tahiti have 4 shader arrays (GCN parlance).

Tahiti's organization of CUs within each shader array is simple because there are 8 CUs per shader array. Four CUs within a shader array share I$ (instruction cache) and K$ (constant cache). Each CU has its own L1$ (level one cache) for textures/buffers. Also, each CU has its own LDS (local data share).

Pitcairn's organization of CUs within each shader array is a bit more complex because there are 5 CUs within each shader array. Within each shader array, 3 CUs share an I$ and K$ and the remaining 2 CUs share an I$ and K$. Each CU has its own L1$ and LDS.

Cape Verde has the same arrangement of Pitcairn except it only has 2 shader arrays.

Hopefully this clarifies things. If this information is not in the OpenCL Programmer's Guide (it should be), I'll make sure it's updated.
 
I only used the terminology you used above. Pitcairn and Tahiti have 4 shader arrays (GCN parlance).
Actually, Dave called this engines (and some docs I looked into, too). I didn't use the word "array" at all. I always assumed array means all the CUs together (that's how it was named in the last documentation I looked into).
Pitcairn's organization of CUs within each shader array is a bit more complex because there are 5 CUs within each shader array. Within each shader array, 3 CUs share an I$ and K$ and the remaining 2 CUs share an I$ and K$. Each CU has its own L1$ and LDS.

Cape Verde has the same arrangement of Pitcairn except it only has 2 shader arrays.

Hopefully this clarifies things. If this information is not in the OpenCL Programmer's Guide (it should be), I'll make sure it's updated.
I was editing my post above with this possibility while you wrote that. But yes, that clarifies things. Thanks for that. It means the information given out during the CV and Pitcairn launches were maybe not fully correct, but now everything makes sense again.
 
Back
Top